Jim,
I think this is an artifact of the demo. The initial images are grayscale,
to get a binary input I had to put a threshold on each pixel value. This
happens in sp_viewer.py in:
def _convertToVector(self, image):
'''
Returns a bit vector representation (list of ints) of a PIL image.
'''
# Convert the image to black and white
image = image.convert('1')
# Pull out the data, turn that into a list, then a numpy array,
# then convert from 0 255 space to binary with a threshold.
# Finnally cast the values into a type CPP likes
vector = (numpy.array(list(image.getdata())) < 100).astype('uint32')
return vector
Change that 100 to something else and see what happens.
Ian
On Sun, Jun 1, 2014 at 4:43 PM, Jim Bridgewater <[email protected]> wrote:
> I tried a different set of images and interestingly the spatial pooler
> groups several of the input patterns even when it is not necessary
> because there are plenty of unused columns available. Increasing the
> amount that permanences are decreased when a synapse does not
> contribute to column activation got each column to converge on a
> representation for a single input image as long as there were enough
> columns available. This makes sense and is shown in the attached
> document.
>
> However, I don't understand why, in this very simple example, the
> permanences do not come to form perfect representations of the input
> images.
>
> For these images the representations do not fill in completely, there
> are synapses that are active when the column is active, but never
> become connected.
>
> Anyone know why this happens?
>
> On Fri, May 30, 2014 at 1:51 AM, <[email protected]> wrote:
> > Send nupic mailing list submissions to
> > [email protected]
> >
> > To subscribe or unsubscribe, visit
> >
> http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
> > or, via email, send a message with subject or body 'help' to
> > [email protected]
> >
> > You can reach the person managing the list at
> > [email protected]
> >
> > When replying, please edit your Subject line so it is more specific
> > than "Re: Contents of nupic digest..."
> >
> >
> > Today's Topics:
> >
> > 1. Re: Some observations from Ian's Fall 2013 SP demo (Ian Danforth)
> > 2. Re: Some observations from Ian's Fall 2013 SP demo
> > (Jim Bridgewater)
> >
> >
> > ----------------------------------------------------------------------
> >
> > Message: 1
> > Date: Thu, 29 May 2014 20:27:51 -0700
> > From: Ian Danforth <[email protected]>
> > To: "NuPIC general mailing list." <[email protected]>
> > Subject: Re: [nupic-discuss] Some observations from Ian's Fall 2013 SP
> > demo
> > Message-ID:
> > <CAOajdBpF8XTt0=
> [email protected]>
> > Content-Type: text/plain; charset="utf-8"
> >
> > Jim,
> >
> > Thanks for writing this up and sharing it with the list! This is both
> > expected and good behavior! Let me see if I can shift your intuitions a
> bit
> > so you'll agree with that :)
> >
> > To start out, what is the SP good for? It does two interesting things, it
> > each column memorizes things it's seen and together they perform
> > dimensionality reduction.
> >
> > Each column is a bit like a film negative, the more times it's exposed
> to a
> > pattern the more it will resemble that pattern. If you have a double
> > exposure, to multiple patterns, it might come to represent both, or a bit
> > of each. If you show one column the same pattern again and again it will
> > form a very strong representation of it. In my demo there are so few
> > patterns each column can come to perfectly represent the input it's
> shown.
> > If you force columns to become active in the presence of more than one of
> > those patterns it will become a representation of a blend of the two. If
> > you make a single column see all the patterns it will blend them all!
> Just
> > like a negative exposed again and again to several scenes. Each column
> is a
> > powerful feature learning system in it's own right, and when they are
> > allowed to compete, they naturally sort themselves into groups of good
> > representations. By forcing them to come on when they would normally be
> > out-competed, they blend into the all-purpose representations you saw.
> >
> > But why doesn't the SP naturally take advantage of the combinatorial
> nature
> > of having multiple columns on? Well that property really only comes into
> > play when you need dimensionality reduction. You need to start with a set
> > of inputs >> than the number of columns. Then when a column is exposed
> over
> > and over, it never sees exactly the same input, and only the common areas
> > to all the inputs (those bits that are most similar) will be retained in
> > the snapshot it retains. This is the pooling of similar, but not exactly
> > similar, features into the activity of each column. Once you have these
> set
> > of features, you *then* use the combinatorial property of having multiple
> > columns on so that you can assemble a complete representation from a set
> of
> > features. If four columns represent vertical and horizontal lines, you
> > might find the combination of all four representing a box. It is the
> > ability to compose complex representations out of a much smaller set of
> > simpler representations where the combinatorial numbers you're talking
> > about come into play. Instead of taking 10,000 columns to learn to
> > differentiate 10,000 things, you can get different representations from a
> > much much smaller set of columns. The unique combination of features in
> > each input providing the eventual unique SDR.
> >
> > I hope this helps in understanding the results! A good next question
> might
> > be, what is the minimum number of inputs you need before having 1 column
> /
> > input becomes unreasonable and you really *need* the property of
> > dimensionality reduction?
> >
> > Ian
> >
> >
> > On Thu, May 29, 2014 at 4:32 PM, Jim Bridgewater <[email protected]>
> wrote:
> >
> >> Hi everyone,
> >>
> >> I'm working on an optical character recognition project for the season
> >> of nupic this summer. I've been playing around with Ian Danforth's
> >> Fall 2013 hackathon demo of the spatial pooler and put together the
> >> attached pdf document which shows some observations of how the spatial
> >> pooler behaves with different numbers of active columns. Maybe some
> >> people on this list have thought about this before and have some
> >> comments about why the SP displays this behavior.
> >>
> >>
> >>
> >> ---------- Forwarded message ----------
> >> From: Scott Purdy <[email protected]>
> >> Date: Thu, May 29, 2014 at 4:02 PM
> >> Subject: Re: Update
> >> To: Jim Bridgewater <[email protected]>
> >>
> >>
> >> The write up is quite nice, thanks for putting that together. You
> >> might send it to the discuss list in case anyone else is curious. I
> >> don't think the results are particularly suprising. In particular, the
> >> SP is known to not always perform well when the number of input bits
> >> or columns is small or there are a small number of input patterns. I
> >> will try to discuss with Subutai tomorrow to confirm that there is
> >> nothing wrong with your experimental method and see if I can get more
> >> details on the onion charts.
> >>
> >> I think the onion charts will make more sense on a more sophisticated
> >> data set as it would likely suffer from the same issues with small
> >> numbers of inputs.
> >>
> >>
> >> On Thu, May 29, 2014 at 1:50 PM, Jim Bridgewater <[email protected]>
> >> wrote:
> >> >
> >> > Hi Scott,
> >> >
> >> > I've been playing with Ian's demo a bit to get a feel for how the
> >> > spatial pooler works. I suppose it was intended to be used with only
> >> > a small % of active columns, but I do find it interesting how poorly
> >> > it performs when 50% of the columns are active since it has the
> >> > greatest number of output values in this case. See attached pdf.
> >> >
> >> > I want to work on a way of showing the results that makes it easier to
> >> > see how the representation of each input image changes as the sp is
> >> > trained and whether the outputs are different for each unique input.
> >> > I was thinking about converting the sp output to a hexadecimal value
> >> > and plotting that vs time for each input image.
> >> >
> >> > Maybe you guys could send me some examples of the onion graphs Subutai
> >> > mentioned so I can see how they work.
> >> >
> >> > --
> >> > Jim Bridgewater, PhD
> >> > Arizona State University
> >> > 480-227-9592
> >>
> >>
> >>
> >>
> >> --
> >> Jim Bridgewater, PhD
> >> Arizona State University
> >> 480-227-9592
> >>
> >> _______________________________________________
> >> nupic mailing list
> >> [email protected]
> >> http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
> >>
> >>
> > -------------- next part --------------
> > An HTML attachment was scrubbed...
> > URL: <
> http://lists.numenta.org/pipermail/nupic_lists.numenta.org/attachments/20140529/030f10df/attachment-0001.html
> >
> >
> > ------------------------------
> >
> > Message: 2
> > Date: Fri, 30 May 2014 01:51:05 -0700
> > From: Jim Bridgewater <[email protected]>
> > To: [email protected]
> > Subject: Re: [nupic-discuss] Some observations from Ian's Fall 2013 SP
> > demo
> > Message-ID:
> > <
> camma+tjpw4es9ovrvy0uf1+stxxtkktpuzqt7jmtzoxxsw8...@mail.gmail.com>
> > Content-Type: text/plain; charset="utf-8"
> >
> > I realized that I was thinking about the number of possible output
> > values incorrectly. An SP that has 16 columns and is configured to
> > have 8 active columns technically has over 12,000 possible output
> > values, but only 2 of those output values have no active columns in
> > common so an SP configured this way can only distinguish between 2
> > unrelated input patterns. In the tests I ran the two patterns are
> > blank and non-blank. I have updated the attached document to reflect
> > this realization.
> >
> > On Thu, May 29, 2014 at 4:32 PM, Jim Bridgewater <[email protected]>
> wrote:
> >> Hi everyone,
> >>
> >> I'm working on an optical character recognition project for the season
> >> of nupic this summer. I've been playing around with Ian Danforth's
> >> Fall 2013 hackathon demo of the spatial pooler and put together the
> >> attached pdf document which shows some observations of how the spatial
> >> pooler behaves with different numbers of active columns. Maybe some
> >> people on this list have thought about this before and have some
> >> comments about why the SP displays this behavior.
> >>
> >>
> >>
> >> ---------- Forwarded message ----------
> >> From: Scott Purdy <[email protected]>
> >> Date: Thu, May 29, 2014 at 4:02 PM
> >> Subject: Re: Update
> >> To: Jim Bridgewater <[email protected]>
> >>
> >>
> >> The write up is quite nice, thanks for putting that together. You
> >> might send it to the discuss list in case anyone else is curious. I
> >> don't think the results are particularly suprising. In particular, the
> >> SP is known to not always perform well when the number of input bits
> >> or columns is small or there are a small number of input patterns. I
> >> will try to discuss with Subutai tomorrow to confirm that there is
> >> nothing wrong with your experimental method and see if I can get more
> >> details on the onion charts.
> >>
> >> I think the onion charts will make more sense on a more sophisticated
> >> data set as it would likely suffer from the same issues with small
> >> numbers of inputs.
> >>
> >>
> >> On Thu, May 29, 2014 at 1:50 PM, Jim Bridgewater <[email protected]>
> wrote:
> >>>
> >>> Hi Scott,
> >>>
> >>> I've been playing with Ian's demo a bit to get a feel for how the
> >>> spatial pooler works. I suppose it was intended to be used with only
> >>> a small % of active columns, but I do find it interesting how poorly
> >>> it performs when 50% of the columns are active since it has the
> >>> greatest number of output values in this case. See attached pdf.
> >>>
> >>> I want to work on a way of showing the results that makes it easier to
> >>> see how the representation of each input image changes as the sp is
> >>> trained and whether the outputs are different for each unique input.
> >>> I was thinking about converting the sp output to a hexadecimal value
> >>> and plotting that vs time for each input image.
> >>>
> >>> Maybe you guys could send me some examples of the onion graphs Subutai
> >>> mentioned so I can see how they work.
> >>>
> >>> --
> >>> Jim Bridgewater, PhD
> >>> Arizona State University
> >>> 480-227-9592
> >>
> >>
> >>
> >>
> >> --
> >> Jim Bridgewater, PhD
> >> Arizona State University
> >> 480-227-9592
> >
> >
> >
> > --
> > Jim Bridgewater, PhD
> > Arizona State University
> > 480-227-9592
> > -------------- next part --------------
> > A non-text attachment was scrubbed...
> > Name: update.pdf
> > Type: application/pdf
> > Size: 274713 bytes
> > Desc: not available
> > URL: <
> http://lists.numenta.org/pipermail/nupic_lists.numenta.org/attachments/20140530/84824396/attachment.pdf
> >
> >
> > ------------------------------
> >
> > Subject: Digest Footer
> >
> > _______________________________________________
> > nupic mailing list
> > [email protected]
> > http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
> >
> >
> > ------------------------------
> >
> > End of nupic Digest, Vol 13, Issue 45
> > *************************************
>
>
>
> --
> Jim Bridgewater, PhD
> Arizona State University
> 480-227-9592
>
> _______________________________________________
> nupic mailing list
> [email protected]
> http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
>
>
_______________________________________________
nupic mailing list
[email protected]
http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
